TUA-Bench and the General-Purpose Terminal Agent: What 120 Non-Coding Tasks Reveal About Codex CLI’s Next Frontier

Every benchmark Codex CLI has conquered so far — SWE-bench, Terminal-Bench, DeepSWE — frames the terminal agent as a coding tool. TUA-Bench, published on 26 June 2026 by Chen et al., reframes the question entirely: can the same agent that patches your Go service also resize a chart in a LibreOffice spreadsheet, draft an email, run an OpenFOAM simulation, and segment a medical image — all from a terminal session?¹

The answer, across five agent frameworks and eight frontier models, is “sometimes” — and the gap between “sometimes” and “reliably” is where the next wave of Codex CLI configuration work lives.

What TUA-Bench Actually Measures

TUA-Bench comprises 120 hand-crafted tasks across five families, each validated by domain experts at PhD level¹:

Task Family	Tasks	Share	Example
Office & Productivity	46	38.3%	Resize a chart in a `.xlsx`, reformat a `.docx` table
Web & Information	22	18.3%	Retrieve live-web data, summarise search results
System & Software Ops	20	16.7%	Configure a service, parse logs, manage packages
Scientific & Engineering	20	16.7%	Run CellProfiler pipelines, OpenFOAM CFD, EnergyPlus
Multimedia & Design	12	13.3%	Edit images, process audio, manipulate slides

Each task runs inside an isolated Linux container with a deterministic setup script. Scoring is execution-based — it checks the final environment state against ground truth, not the trajectory the agent took to get there¹. Each task has a 2,400-second time limit. Pass@1, pass@5, and all-5 (consistent success across five attempts) metrics are reported.

How Codex CLI Performed

Five agent frameworks were evaluated, each paired with multiple models¹:

graph LR
    subgraph Agents
        A[Claude Code]
        B[Codex CLI]
        C[OpenHands]
        D[Mini-SWE-Agent]
        E[Terminus-2]
    end
    subgraph Models
        F[Claude Opus 4.8]
        G[GPT-5.5]
        H[Gemini 3.1 Pro]
        I[GPT-5.4 mini]
        J[Claude Haiku 4.5]
    end
    A --> F
    B --> G
    C --> F
    D --> G
    E --> G

The top configurations on pass@1:

Agent + Model	Score
Claude Code + Opus 4.8 (max reasoning)	65.8%
Codex CLI + GPT-5.5	64.7%
OpenHands + Opus 4.8	63.4%
Mini-SWE-Agent + GPT-5.5	62.4%
Terminus-2 + GPT-5.5	60.1%

The 1.1 percentage point gap between Claude Code and Codex CLI is within noise, but three findings matter for Codex CLI practitioners.

Three Findings That Matter

1. The Harness Effect Persists Beyond Coding

TUA-Bench confirms what PERFOPT-Bench and Terminal-Bench established for coding tasks: the choice of harness can have a comparable effect to the choice of underlying model¹. Claude Opus 4.8 achieved 63.4% with OpenHands but only 59.7% with Terminus-2 — a 3.7pp swing from scaffold alone. GPT-5.5 showed a 4.6pp spread across the three frameworks it was tested with.

For Codex CLI users, this means your config.toml settings, AGENTS.md instructions, MCP server configuration, and sandbox mode matter as much for non-coding tasks as they do for SWE-bench. The agent harness is the product.

2. Office & Productivity Is the Hardest Family

Despite being the most “routine” category, Office & Productivity tasks showed the widest variance across agents. Chart resizing, slide layout modification, and spreadsheet formula manipulation in headless LibreOffice consistently defeated all configurations¹. Several tasks appeared as “near-horizontal red bands” in the results matrix — persistent failures that no model or harness could solve.

This matters because Codex CLI’s 8 million users now include roughly 20% non-developers². These knowledge workers are precisely the users most likely to attempt document-editing workflows through the terminal. TUA-Bench suggests the tooling gap here is real: without proper MCP servers or skills wrapping LibreOffice CLI operations, even frontier models struggle.

3. Reasoning Effort Hits Diminishing Returns

Increasing Claude Opus 4.8 from high to extreme (max) reasoning effort added only 2.3 percentage points while nearly doubling token costs¹. For Codex CLI, this maps directly to the reasoning_effort configuration key and the Ultra mode toggle. The TUA-Bench evidence suggests that for general-purpose terminal tasks, the high setting delivers the best cost-performance ratio — reserving max for scientific/engineering tasks where the marginal gain justifies the spend.

Configuring Codex CLI for General-Purpose Terminal Work

TUA-Bench tasks demand capabilities that a default Codex CLI installation lacks. Here is a configuration pattern for general-purpose terminal workflows:

config.toml: Sandbox and Network Access

Most TUA-Bench tasks require network access (web research, email, package installation) and filesystem writes outside the project directory:

[sandbox]
mode = "danger-full-access"   # Required for system-level tasks

[network]
network_proxy_allowed_domains = [
  "*.google.com",
  "*.wikipedia.org",
  "api.openweathermap.org",
]

⚠️ danger-full-access should only be used in disposable container environments, never on a development machine with production credentials. For TUA-Bench-style workflows on real machines, use workspace-write with explicit approval for system-level operations.

AGENTS.md: Task-Family Routing

An AGENTS.md file can encode task-family awareness:

## General-Purpose Terminal Agent

### Document Editing
- Use `libreoffice --headless --calc` for spreadsheet operations
- Use `libreoffice --headless --writer` for document manipulation
- Always verify output with `libreoffice --headless --convert-to pdf` and check page count
- Prefer python-docx / openpyxl over raw LibreOffice CLI when the task permits

### Web Research
- Use `curl` with `-L` for redirects and `jq` for JSON APIs
- Prefer structured APIs over scraping when available
- Cache responses to avoid repeated network calls

### Scientific Workflows
- Check that simulation tools (OpenFOAM, CellProfiler, EnergyPlus) are installed before attempting tasks
- Parse output files programmatically rather than visually inspecting them

MCP Servers for Non-Coding Tasks

The TUA-Bench results highlight a tooling gap that MCP servers can fill. Several community MCP servers already exist for document workflows³:

[mcp_servers.docx-cli]
command = "npx"
args = ["docx-cli", "serve"]

[mcp_servers.xlsx-tools]
command = "npx"
args = ["xlsx-mcp-server"]

flowchart TD
    A[User Task] --> B{Task Family?}
    B -->|Office| C[docx-cli / xlsx MCP]
    B -->|Web| D[curl + jq / Playwright MCP]
    B -->|System| E[Shell commands / package managers]
    B -->|Scientific| F[Domain-specific CLI tools]
    B -->|Multimedia| G[ImageMagick / ffmpeg]
    C --> H[LibreOffice headless verification]
    D --> H2[Response caching]
    E --> H3[Approval hooks for destructive ops]
    F --> H4[Output file parsing]
    G --> H5[Format conversion verification]

Named Profiles for Task Families

Codex CLI’s named profiles (available since v0.142) let you switch configurations per task type without editing config.toml:

[profiles.office]
model = "gpt-5.6-luna"
reasoning_effort = "medium"
sandbox_mode = "workspace-write"

[profiles.scientific]
model = "gpt-5.6-terra"
reasoning_effort = "high"
sandbox_mode = "danger-full-access"

[profiles.coding]
model = "gpt-5.6-terra"
reasoning_effort = "high"
sandbox_mode = "workspace-write"

Then invoke with:

codex --profile scientific "Run the OpenFOAM mesh generation and report convergence"
codex --profile office "Resize the chart in Q3-report.xlsx to 800x600 and bold the axis labels"

The Broader Signal: Terminal Agents Are Becoming General Agents

TUA-Bench sits alongside TerminalWorld (1,530 tasks from 80,870 terminal recordings)⁴ and Terminal-Bench 2.0 (89 tasks, ICLR 2026)⁵ in a growing family of benchmarks that evaluate terminal agents as general-purpose computer-use agents, not just coding assistants.

The trajectory is clear:

timeline
    title Terminal Agent Benchmark Evolution
    2024 : SWE-bench (coding only)
    2025 : Terminal-Bench 1.0 (CLI tasks)
    Jan 2026 : Terminal-Bench 2.0 (89 hard CLI tasks)
    May 2026 : TerminalWorld (1530 real-world tasks)
    Jun 2026 : TUA-Bench (120 general-purpose tasks including office and scientific)

For Codex CLI developers, the implication is that your AGENTS.md, config.toml, MCP servers, and skills are not just coding infrastructure — they are the scaffolding for a general-purpose terminal agent. The 20% of Codex users who are non-developers² are the leading indicator. The benchmark frontier has followed them.

What to Watch

Three developments will determine whether Codex CLI closes the TUA-Bench gap:

LibreOffice MCP server maturity. Office & Productivity tasks are the weakest family. A robust LibreOffice MCP server that exposes chart manipulation, slide layout, and formula editing as structured tools would shift scores significantly.
GPT-5.6 Terra on TUA-Bench. The benchmark tested GPT-5.5; Terra’s improved long-context performance (+55.5pp on Graphwalks)⁶ could help with the multi-step scientific workflows where context accumulates rapidly.
Agent Plugins for domain tools. Codex CLI v0.146 introduced Agent Plugin manifests⁷. Scientific task families need plugins that wrap CellProfiler, OpenFOAM, and EnergyPlus CLI interfaces with structured input/output schemas.

Citations

Chen, S., Wang, L., Yang, X., Liu, Z., Cong, Y., Ji, Y., Zhou, F., Zhang, X., Yang, F., & Zeng, B. (2026). “TUA-Bench: A Benchmark for General-Purpose Terminal-Use Agents.” arXiv:2606.28480. https://arxiv.org/abs/2606.28480 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷
OpenAI. (2026). “Codex Passes 5 Million Weekly Users.” Reported by Tech Insider Ireland, June 2026. https://tech-insider.org/ie/openai-codex-5-million-users-2026/ ↩ ↩²
docx-cli: MIT-licensed CLI and agent skill for Word .docx files. https://vibecodinghub.org/tools/docx-cli ↩
TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks. arXiv:2605.22535. https://arxiv.org/abs/2605.22535 ↩
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces. arXiv:2601.11868, ICLR 2026. https://arxiv.org/abs/2601.11868 ↩
OpenAI. (2026). “GPT-5.4 Retirement and GPT-5.6 Terra/Luna Migration.” Announced 31 July 2026. https://developers.openai.com/codex/changelog ↩
OpenAI. (2026). Codex CLI v0.146.0 release notes, 29 July 2026. https://github.com/openai/codex/releases ↩